Coupled Tensor-Tensor Completion Method with Applications in Drug Repurposing
Maryam Bagherian, Albert Hung, Ivo Dinov, Joshua Welch
math.NA · cs.LG
Quick verdict
CTTC extends tensor completion to tensor-valued side information with big RSE gains on DTD/LINCS; compelling, though the runtime claim conflicts with Table 2.
Summary
The paper introduces Coupled Tensor-Tensor Completion (CTTC), a penalized optimization framework for completing a partially observed tensor while exploiting side information given as tensors — multiple stacked similarity arrays per mode — rather than a single matrix per mode as in prior coupled methods. Each mode carries learned Mahalanobis-type metric matrices U(ℓ), constrained to the special linear group, and the objective couples the completed tensor to all matricizations of the similarity tensors via trace terms. The authors derive an alternating minimization scheme, prove well-definedness and stationarity of accumulation points under standard assumptions (following Bertsekas), and give a perturbation bound for the tensor Mahalanobis distance. On two public drug-repurposing datasets — DTD (810×302×542) and LINCS (6780×70×1104) — CTTC reports much lower RSE than HaLRTC, CTRC, Cell, and NTD-DR across 10–50% missing rates (e.g., DTD at 10%: 0.0038 vs 0.0401 for the best baseline CTRC), and an ablation (Table 4) shows DTD RSE at 20% missing rises from 0.0154 to 0.4190 when all side information is switched off. Stratified evaluations over chemical classes and gene families, plus unseen-drug/gene holdouts, are also reported. Main limitations: the abstract's runtime-superiority claim is contradicted by Table 2 (HaLRTC is faster on both datasets), regularization parameters appear tuned during execution, raising leakage concerns; coupled-nuclear-norm baselines are cited but not compared; and code is only promised upon publication.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CTTC is a tensor completion framework that accepts tensor-valued (multi-layer) side information per mode, generalizing coupled matrix-matrix and coupled tensor-matrix completion.
- An alternating minimization algorithm solves the CTTC objective and converges to a stationary point under standard assumptions.
- CTTC achieves substantially lower RSE than HaLRTC, CTRC, Cell, and NTD-DR on DTD and LINCS across 10–50% missing rates (e.g., 0.0038 vs 0.0401 on DTD at 10% missing).
- Side information drives the gains: setting all ϱℓ = 0 raises DTD RSE from 0.0154 to 0.4190 at 20% missing (Table 4).
- CTTC can predict effects for entirely unseen drugs or genes, though not for unseen cell lines due to sparse cell-line side information.
Strengths
Novel problem formulation
Extends coupled completion from matrix-valued to tensor-valued side information (§1, Fig. 1): each mode can carry arbitrarily many stacked similarity layers (eq. 8), generalizing the authors' prior coupled matrix-tensor line of work [5,7].
Large, consistent empirical margins
Table 1: on DTD at 10% missing, CTTC RSE 0.0038 (fit 0.99) vs 0.0401 for the best baseline CTRC; on LINCS at 50% missing, 0.0179 vs 0.2245 for NTD-DR; CTTC wins at every missing rate on both datasets over 10 repetitions with reported SDs.
Clean ablation of the core idea
Theory beyond black-box use
Above-average evaluation breadth and honesty
Weaknesses
Abstract run-time claim contradicted by own data
Tuning procedure invites optimistic bias
Cited coupled side-information baselines never compared
Theory largely imported; key assumptions unverified
Presentation and internal-consistency issues
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.03190, math.NA/cs.LG) proposes Coupled Tensor-Tensor Completion (CTTC), an extension of low-rank tensor completion that ingests side information in tensor form — multiple stacked similarity arrays per mode — rather than a single matrix per mode, as in prior coupled methods. Central claims: (i) novelty of tensor-valued side information; (ii) an alternating algorithm with convergence to a stationary point; (iii) large accuracy gains over HaLRTC, CTRC, Cell, and NTD-DR on DTD and LINCS; (iv) runtime superiority. Claim (iv) is contradicted by the paper's own Table 2 and must be corrected; claims (i)–(iii) are supported within the limits discussed below.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Coupled Tensor-Tensor Completion (CTTC)?
CTTC is a method for filling in missing values of a large multi-dimensional table (a tensor), such as drug–gene–cell-line effect data, by also exploiting side information supplied as tensors — multiple stacked similarity arrays per dimension — instead of the single similarity matrix per dimension used by earlier coupled completion methods. It learns a distance metric per dimension and optimizes all components by alternating minimization until convergence.
How much better is CTTC than existing tensor completion methods?
On the paper's two public benchmarks, CTTC reports much lower relative squared error (RSE) than HaLRTC, CTRC, Cell, and NTD-DR at 10–50% missing data — for example 0.0038 vs 0.0401 (best baseline CTRC) on the DTD drug-target-disease dataset at 10% missing, and 0.0179 vs 0.2245 (NTD-DR) on LINCS at 50% missing. Its runtime is slower than HaLRTC but faster than the other three baselines.
Can CTTC predict drug effects for new drugs never seen during training?
Yes, according to the paper's case study (Table 3): when all held-out entries are drawn from the drug set or the gene set, RSE stays comparable to random holdout (e.g., 0.103 for drugs at 10% missing), suggesting useful prediction for unseen drugs and genes. However, prediction for unseen cell lines did not produce meaningful results, which the authors attribute to sparse cell-line similarity information.
Is the CTTC code and data available?
Not at the arXiv stage. The paper states that a subset of processed data, code (a MATLAB implementation), and usage instructions will be posted to a GitHub repository (github.com/mbagherian/CTTC) upon publication. All input datasets (LINCS, ChemicalChecker, DrugBank, NTD-DR, STRING, RegNetwork) are described as publicly available.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure image contents (Figs. 1–4) are not recoverable from the text extraction; assessments rely on captions and in-text descriptions only.
- Table 1 is extracted with fit/RSE pairs interleaved in running order; method-to-value alignment was cross-checked against the header order and the narrative claims, but a column shift cannot be fully ruled out.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.